feat(quic): heap-sampling soak test, FD-leak canary, phantom-stream guard, loss harness

Follow-up to b8c6e080 addressing the gaps I called out in the
"is this the best we can do?" reply.

1. **Heap-sampling soak test** (`QuicHeapSoakTest`). Long-form, default-
   skipped via `-PquicSoakSeconds=N` propagated by `quic/build.gradle.kts`
   to the jvmTest task. Without the property the test early-returns with
   a printed SKIPPED line, so `./gradlew test` stays fast for CI. With
   the property, drives moq-lite-shaped peer-uni churn at ~50 streams/s
   for N seconds, samples `totalMemory - freeMemory` six times across
   the run, and fails if the post-warmup → final delta exceeds 10 MB
   (the acceptance threshold from the audio-rooms soak prompt).
   Production use: `-PquicSoakSeconds=1800` for the 30-minute soak.

2. **FD-leak canary** added to `QuicConnectionDriverLifecycleTest`. On
   Linux, samples `/proc/self/fd` size before / after the 100-session
   loop; banded at +16 entries for ambient JVM noise. macOS / Windows
   silently no-op because /proc isn't there. Catches socket / pipe
   leaks the thread-count check would miss.

3. **Phantom-stream guard.** Added `retiredStreamIdSet` (capped FIFO
   ring at 4 096 entries, ~80 s of moq-lite churn) plus
   `isStreamIdRetiredLocked` on the connection. Parser checks before
   `getOrCreatePeerStreamLocked` and drops STREAM frames the peer
   retransmits on already-retired streams. Eliminates the
   "duplicate ACK lost → peer retransmits FIN → we mint a phantom
   QuicStream" edge case I papered over in the previous commit.
   Pinned by `phantomGuardDropsRetransmitOnRetiredPeerStream`.

4. **moq-lite loss harness** (`MoqLiteLossHarnessTest`). First pass at
   soak target #5: drive 50 best-effort group streams with 5%
   uniform packet loss, assert the listener surfaces ≥ 90% with
   payloads intact and the connection stays CONNECTED. Out of scope
   here: reorder injection, latency-under-loss measurement, full
   end-to-end with a real moq-lite publisher.

Tests:
 - `QuicHeapSoakTest` — gated, validates 10MB heap acceptance band.
 - `QuicConnectionDriverLifecycleTest::repeatedSessionLifecycleDoesNotLeakThreads`
   — now also enforces /proc/self/fd bound.
 - `StreamRetirementSoakTest::phantomGuardDropsRetransmitOnRetiredPeerStream`
   — pins the duplicate-frame drop semantics.
 - `MoqLiteLossHarnessTest` — 2 cases (lossy + lossRate=0 baseline).

https://claude.ai/code/session_018KPKWRg5baX5Anf7zfEyec
This commit is contained in:
Claude
2026-05-07 22:10:27 +00:00
parent bfc983bff7
commit c65eef6927
7 changed files with 706 additions and 7 deletions
+11
View File
@@ -124,3 +124,14 @@ tasks.register<JavaExec>("interop") {
args(host, port)
systemProperty("interopTimeoutSec", timeoutSec)
}
// Long-form audio-rooms soak test. Disabled by default — `./gradlew test`
// must stay fast for CI. Opt in with `-PquicSoakSeconds=N` (e.g. 1800 for
// the 30-minute run from the soak prompt). Without the property the test
// class checks for null and skips via `Assume.assumeTrue`.
tasks.withType<Test>().configureEach {
val soakSeconds = (project.findProperty("quicSoakSeconds") as? String)
if (soakSeconds != null) {
systemProperty("quicSoakSeconds", soakSeconds)
}
}
@@ -350,6 +350,33 @@ class QuicConnection(
internal var retiredStreamsCount: Long = 0L
private set
/**
* FIFO ring of recently-retired stream IDs. The parser consults
* this in [getOrCreatePeerStreamLocked] to drop duplicate STREAM
* frames the peer might retransmit on a stream we've already torn
* down — without this guard, a retransmit (which can happen if our
* ACK of the FIN frame was lost and the peer's loss detector
* re-fired) would create a fresh QuicStream object, deliver
* duplicate bytes to the application, and bump
* [peerInitiatedUniCount] a second time.
*
* Bounded at [RETIRED_STREAM_ID_RING_SIZE] entries. Eviction is
* FIFO so a long-running session never grows this set unbounded —
* at moq-lite churn rates of ~50 streams/sec, the ring covers the
* last ~80 seconds of retired IDs, far longer than the peer's
* loss-detection retransmit window (a small multiple of RTT).
*
* Out-of-bounds duplicate retransmits (extremely rare — would need
* the peer's ACK to be lost AND our retransmit not arriving for
* many seconds) fall through to the existing
* "create-and-immediately-retire" path, which is the previous
* pre-guard behavior and merely costs a re-iteration.
*
* Caller of any read/write must hold [streamsLock].
*/
private val retiredStreamIdsOrder = ArrayDeque<Long>()
private val retiredStreamIdSet = HashSet<Long>()
/**
* Peer-advertised concurrent bidirectional stream cap. Initialised from
* [TransportParameters.initialMaxStreamsBidi] when peer params arrive,
@@ -1753,13 +1780,11 @@ class QuicConnection(
* will not re-send anything on the stream.
* - RECEIVE side waits for the parser to have pushed every byte to
* the application's incoming Channel ([ReceiveBuffer.isFullyRead]).
* Subsequent retransmits from the peer (rare — peer only retransmits
* if its own loss-detection fires) would land on a fresh stream
* object created by [getOrCreatePeerStreamLocked], deliver
* duplicate bytes, and re-fire FIN to a closed channel; the moq-lite
* listener treats duplicates as no-ops, so the failure mode is "a
* bit of wasted CPU" rather than "wrong audio". The alternative
* (retain every stream forever) is a hard memory leak in soak.
* Retired stream IDs are recorded in [retiredStreamIdSet] so a
* duplicate STREAM frame the peer retransmits (because its own
* loss-detection fired before our ACK arrived) is dropped at
* [getOrCreatePeerStreamLocked] rather than minting a phantom
* stream that delivers duplicate bytes to the application.
*/
internal fun retireFullyDoneStreamsLocked(): Int {
if (streamsList.isEmpty()) return 0
@@ -1778,6 +1803,12 @@ class QuicConnection(
// resetAcked never having received any STREAM frame at all).
// closeIncoming is idempotent.
stream.closeIncoming()
// Record the id so a peer retransmit on this stream gets
// dropped instead of recreating a phantom stream. FIFO
// ring with bounded size — ancient retired IDs eventually
// age out, but the eviction window is far larger than the
// peer's plausible retransmit horizon.
recordRetiredStreamIdLocked(stream.streamId)
streams.remove(stream.streamId)
it.remove()
removed++
@@ -1794,6 +1825,38 @@ class QuicConnection(
return removed
}
/**
* Add [streamId] to the retired-IDs ring. FIFO eviction once the
* ring is at [RETIRED_STREAM_ID_RING_SIZE] entries — the oldest
* entry is removed from both the ordered deque and the lookup
* set in lockstep. Idempotent on duplicate adds (the
* [retireFullyDoneStreamsLocked] caller already drops the stream
* from `streams` before this fires, so a duplicate add can only
* happen if a caller manually re-records — defensive `add` skip
* keeps the ring entries unique without the cost of a fresh
* dedup pass).
*
* Caller must hold [streamsLock].
*/
private fun recordRetiredStreamIdLocked(streamId: Long) {
if (retiredStreamIdSet.add(streamId)) {
retiredStreamIdsOrder.addLast(streamId)
while (retiredStreamIdsOrder.size > RETIRED_STREAM_ID_RING_SIZE) {
val evicted = retiredStreamIdsOrder.removeFirst()
retiredStreamIdSet.remove(evicted)
}
}
}
/**
* True if [streamId] has been retired *and* is still inside the
* ring's eviction window. Used by [QuicConnectionParser] to drop
* STREAM frames the peer retransmits on already-retired streams.
*
* Caller must hold [streamsLock].
*/
internal fun isStreamIdRetiredLocked(streamId: Long): Boolean = streamId in retiredStreamIdSet
/** Caller must hold [lock]. Pending datagram queue for the driver's send loop. */
internal fun pendingDatagramsLocked(): ArrayDeque<ByteArray> = pendingDatagrams
@@ -1983,6 +2046,24 @@ class QuicConnection(
* audio/video, fresh frames matter more than stale ones.
*/
const val MAX_INCOMING_DATAGRAM_QUEUE: Int = 256
/**
* Capacity of the retired-stream-IDs ring used by
* [recordRetiredStreamIdLocked] / [isStreamIdRetiredLocked].
* Sized so the ring covers ≥ 80 seconds of retirement at
* moq-lite's ~50 streams/sec churn rate, far longer than any
* plausible peer retransmit window (a small multiple of RTT
* — at the absolute worst tens of seconds on lossy mobile
* networks). Older retired IDs eviction-fall-through to the
* existing create-and-immediately-retire path, which is
* functionally correct, just with one extra round of work
* per duplicate.
*
* Memory: 4 096 × (8 bytes Long key + ~32 bytes HashSet
* overhead + 8 bytes ArrayDeque slot) ≈ 200 KB. Trivial vs
* the per-stream object size we're saving by retiring.
*/
const val RETIRED_STREAM_ID_RING_SIZE: Int = 4_096
}
}
@@ -513,6 +513,24 @@ private fun dispatchFrames(
)
return
}
// Phantom-stream guard: drop STREAM frames the peer
// retransmitted on a stream we've already retired. Without
// this, the next line would mint a fresh QuicStream object
// and re-deliver duplicate bytes to the application
// (worse, on a SERVER_UNI stream it would also bump
// peerInitiatedUniCount and trigger a spurious
// MAX_STREAMS_UNI emission). The frame is still
// ack-eliciting — we set ackEliciting above — so our
// ACK goes out and the peer's loss-detector backs off.
// Stays inside the StreamFrame handler so other frame
// types (RESET_STREAM, MAX_STREAM_DATA, STOP_SENDING)
// on retired ids gracefully fall through their existing
// streamByIdLocked == null branches as no-ops.
if (conn.isStreamIdRetiredLocked(frame.streamId) &&
conn.streamByIdLocked(frame.streamId) == null
) {
continue
}
val stream = conn.getOrCreatePeerStreamLocked(frame.streamId)
// RFC 9000 §4.1: peer MUST NOT send beyond the limit we advertised.
// The connection-level kill protects against unbounded memory
@@ -0,0 +1,230 @@
/*
* Copyright (c) 2025 Vitor Pamplona
*
* Permission is hereby granted, free of charge, to any person obtaining a copy of
* this software and associated documentation files (the "Software"), to deal in
* the Software without restriction, including without limitation the rights to use,
* copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the
* Software, and to permit persons to whom the Software is furnished to do so,
* subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
* FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
* COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
* AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
* WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
package com.vitorpamplona.quic.connection
import com.vitorpamplona.quic.frame.StreamFrame
import com.vitorpamplona.quic.stream.StreamId
import com.vitorpamplona.quic.tls.InProcessTlsServer
import com.vitorpamplona.quic.tls.PermissiveCertificateValidator
import kotlinx.coroutines.flow.toList
import kotlinx.coroutines.runBlocking
import kotlinx.coroutines.withTimeoutOrNull
import kotlin.random.Random
import kotlin.test.Test
import kotlin.test.assertEquals
import kotlin.test.assertTrue
/**
* First pass at soak target #5 (transient packet loss + jitter) for
* the moq-lite group-stream shape.
*
* **Scope of this draft.** moq-lite's group streams are best-effort
* (the [SendBuffer.bestEffort] flag drops lost ranges instead of
* retransmitting). On the LISTENER side that means: a peer-uni
* stream that the server originated, whose carrying datagram never
* arrives, is gone the relay does not retransmit it. The listener
* application is expected to observe a small fraction of frames
* "missing" but the connection itself stays healthy and audio for
* the surviving frames continues.
*
* This test pins exactly that contract:
* 1. Server publishes 50 group streams (one StreamFrame + FIN per
* stream, one stream per datagram the moq-lite group shape
* 1:1).
* 2. A configurable loss model drops 5% (uniformly random) of
* server datagrams before [feedDatagram].
* 3. We assert the listener surfaces 90% of the published
* streams ( 45 of 50 at p_loss=0.05) and the connection stays
* CONNECTED.
*
* What's intentionally OUT of scope here:
* - **Reordering / jitter.** moq-lite tolerates reordering by
* construction (each group is its own stream, contiguous within
* itself). A reorder-injecting wrapper is the obvious next
* layer; this draft pins the loss-only contract first.
* - **Reliable bidi STREAM frame retransmit on loss.** Covered by
* the existing CryptoRetransmitTest / MultiplexingRoundTripTest;
* the transferloss / handshakeloss interop testcases drive that
* end-to-end.
* - **Latency under loss.** Worth measuring once we wire a real
* moq-lite publisher; for now `runBlocking` + in-memory pipe
* means "real time" is meaningless.
*
* Production link to chase next:
* - `nestsClient/.../moq/lite/Subscriber.kt` the listener path.
* Verify it doesn't tear down its own subscription on a missing
* group sequence; surfacing the gap to the audio decoder
* (silence frames) is the right behaviour.
*/
class MoqLiteLossHarnessTest {
@Test
fun listenerToleratesFivePercentLossOnGroupStreams() =
runBlocking {
// Deterministic RNG so a CI flake is reproducible: the
// exact dropped indices are seeded from a fixed value.
// Bumping the seed to find pathological loss patterns is
// a good follow-up, but for now we just need any seed
// that exercises the dropped-and-recovered path.
val rng = Random(0xC01DBEEFL)
val groupCount = 50
val lossRate = 0.05
val (client, pipe) = newConnectedClient()
val firstId = StreamId.build(StreamId.Kind.SERVER_UNI, 0L)
val streamIds = (0 until groupCount).map { firstId + 4L * it }
val payloads = streamIds.associateWith { id -> "frame-$id".encodeToByteArray() }
val droppedIds = mutableSetOf<Long>()
for (id in streamIds) {
val drop = rng.nextDouble() < lossRate
if (drop) {
droppedIds += id
continue // Don't even build the packet — same shape as kernel
// dropping the datagram before it reaches the QUIC parser.
}
val frame =
StreamFrame(
streamId = id,
offset = 0L,
data = payloads[id]!!,
fin = true,
)
val packet = pipe.buildServerApplicationDatagram(listOf(frame))!!
feedDatagram(client, packet, nowMillis = 0L)
}
// Drain the listener's incoming Flow for every stream we
// EXPECTED to receive (= all server-uni IDs minus the dropped
// ones). Streams that were dropped never had a QuicStream
// created on our side, so streamById(id) returns null — we
// don't even try to drain those.
val received = mutableMapOf<Long, ByteArray>()
for (id in streamIds) {
if (id in droppedIds) continue
val stream = client.streamById(id) ?: continue
val chunks = withTimeoutOrNull(2_000L) { stream.incoming.toList() } ?: continue
val joined = ByteArray(chunks.sumOf { it.size })
var p = 0
for (c in chunks) {
c.copyInto(joined, p)
p += c.size
}
received[id] = joined
}
// Surviving streams must surface their full payload.
for ((id, bytes) in received) {
assertEquals(
payloads[id]!!.decodeToString(),
bytes.decodeToString(),
"stream $id payload corrupted under loss — best-effort path " +
"must NOT split or drop bytes WITHIN a delivered group",
)
}
// ≥ 90% delivered (lossRate=5% — band leaves room for RNG
// tail-end where the seed happens to drop more than the
// mean).
val expectedFloor = (groupCount * 0.9).toInt()
assertTrue(
received.size >= expectedFloor,
"listener received only ${received.size} / $groupCount groups under " +
"${(lossRate * 100).toInt()}% loss — floor was $expectedFloor. " +
"dropped=${droppedIds.size} ids=${droppedIds.sorted()}",
)
assertEquals(
QuicConnection.Status.CONNECTED,
client.status,
"connection must stay CONNECTED across packet loss on best-effort streams",
)
}
@Test
fun listenerSurfacesEveryFrameWhenLossRateIsZero() =
runBlocking {
// Sanity for the harness itself: with lossRate=0 we must
// see all 50 frames intact. If this fails the loss-shape
// assertions in the lossy test are unreliable.
val (client, pipe) = newConnectedClient()
val groupCount = 50
val firstId = StreamId.build(StreamId.Kind.SERVER_UNI, 0L)
for (i in 0 until groupCount) {
val id = firstId + 4L * i
val payload = "frame-$id".encodeToByteArray()
val frame = StreamFrame(streamId = id, offset = 0L, data = payload, fin = true)
val packet = pipe.buildServerApplicationDatagram(listOf(frame))!!
feedDatagram(client, packet, nowMillis = 0L)
}
var delivered = 0
for (i in 0 until groupCount) {
val id = firstId + 4L * i
val stream = client.streamById(id)!!
val chunks = withTimeoutOrNull(2_000L) { stream.incoming.toList() }
if (chunks != null) delivered += 1
}
assertEquals(groupCount, delivered, "lossRate=0 baseline must deliver every frame")
}
private fun newConnectedClient(): Pair<QuicConnection, InMemoryQuicPipe> =
runBlocking {
val client =
QuicConnection(
serverName = "loss.test",
config =
QuicConnectionConfig(
initialMaxStreamsBidi = 1024,
initialMaxStreamsUni = 1024,
initialMaxData = 16L * 1024 * 1024,
initialMaxStreamDataBidiLocal = 64L * 1024,
initialMaxStreamDataBidiRemote = 64L * 1024,
initialMaxStreamDataUni = 64L * 1024,
),
tlsCertificateValidator = PermissiveCertificateValidator(),
)
val serverScid = ConnectionId.random(8)
val tlsServer =
InProcessTlsServer(
transportParameters =
TransportParameters(
initialMaxData = 16L * 1024 * 1024,
initialMaxStreamDataBidiLocal = 64L * 1024,
initialMaxStreamDataBidiRemote = 64L * 1024,
initialMaxStreamDataUni = 64L * 1024,
initialMaxStreamsBidi = 1024,
initialMaxStreamsUni = 1024,
initialSourceConnectionId = serverScid.bytes,
originalDestinationConnectionId = client.destinationConnectionId.bytes,
).encode(),
)
val pipe =
InMemoryQuicPipe(
client = client,
initialDcid = client.destinationConnectionId.bytes,
serverScid = serverScid,
tlsServer = tlsServer,
)
client.start()
pipe.drive(maxRounds = 16)
assertEquals(QuicConnection.Status.CONNECTED, client.status)
client to pipe
}
}
@@ -207,6 +207,71 @@ class StreamRetirementSoakTest {
)
}
@Test
fun phantomGuardDropsRetransmitOnRetiredPeerStream() =
runBlocking {
// Phantom-stream guard regression: after we've retired a
// peer-uni stream, a duplicate STREAM frame the peer
// retransmits (because its loss-detector fired before our
// ACK reached it) MUST NOT mint a fresh QuicStream object.
// Pre-guard, the duplicate frame would:
// - re-populate `streams[id]` and `streamsList`
// - bump `peerInitiatedUniCount` a second time
// - re-fire `newPeerStreams.addLast(...)`, signalling
// a phantom stream-arrival to application code
// - re-deliver the same bytes on a fresh incoming Flow
// The guard at the parser drops the frame (still
// ack-eliciting, so our ACK still goes out and the peer's
// loss-detector quiets down).
val (client, pipe) = newConnectedClient()
val streamId = StreamId.build(StreamId.Kind.SERVER_UNI, 0L)
val payload = "phantom-test".encodeToByteArray()
val frame = StreamFrame(streamId = streamId, offset = 0L, data = payload, fin = true)
val firstPacket = pipe.buildServerApplicationDatagram(listOf(frame))!!
feedDatagram(client, firstPacket, nowMillis = 0L)
// Drain the original delivery to retire the stream.
client.streamById(streamId)?.incoming?.toList()
drainAll(client, pipe)
assertEquals(
1L,
client.retiredStreamsCount,
"first delivery + drain must retire exactly one stream",
)
assertTrue(
client.isStreamIdRetiredLocked(streamId),
"retired-id ring must hold the just-retired stream id",
)
val countAfterFirstDelivery = client.peerInitiatedUniCount
// Replay the same STREAM+FIN frame — i.e. peer retransmit.
val replayPacket = pipe.buildServerApplicationDatagram(listOf(frame))!!
feedDatagram(client, replayPacket, nowMillis = 0L)
drainAll(client, pipe)
// No phantom stream — counters must not move.
assertEquals(
1L,
client.retiredStreamsCount,
"duplicate STREAM frame must not create a phantom stream that re-retires",
)
assertEquals(
countAfterFirstDelivery,
client.peerInitiatedUniCount,
"peerInitiatedUniCount must not double-count the retransmitted FIN",
)
assertEquals(
0,
client.streamsListLocked().size,
"streamsList must stay empty after the phantom-guard drop",
)
assertEquals(
QuicConnection.Status.CONNECTED,
client.status,
"connection must stay CONNECTED across a duplicate STREAM frame",
)
}
@Test
fun retirementPreservesConnectionLevelMaxDataAccounting() =
runBlocking {
@@ -29,6 +29,7 @@ import kotlinx.coroutines.cancel
import kotlinx.coroutines.delay
import kotlinx.coroutines.runBlocking
import kotlinx.coroutines.withTimeoutOrNull
import java.io.File
import java.net.DatagramSocket
import java.net.InetSocketAddress
import kotlin.test.AfterTest
@@ -139,6 +140,7 @@ class QuicConnectionDriverLifecycleTest {
// shows up as a leak.
warmUpDispatchersIo()
val baseline = liveThreadCount()
val baselineFds = liveFileDescriptorCount()
val sessions = 100
for (i in 0 until sessions) {
val parent = CoroutineScope(SupervisorJob() + Dispatchers.IO)
@@ -182,6 +184,23 @@ class QuicConnectionDriverLifecycleTest {
"thread count grew by $growth across $sessions sessions " +
"(baseline=$baseline final=$finalCount). Anything > 16 indicates a leak.",
)
// FD-leak canary. On Linux, /proc/self/fd holds one entry per
// open file descriptor (sockets + pipes + regular files).
// A leak that misses socket.close() would show up here as
// ~1 FD per session — 100 sessions → growth ≥ 100 with
// certainty. We band at 16 (same headroom as threads). On
// platforms without /proc, [liveFileDescriptorCount] returns
// -1 and this branch silently no-ops.
val finalFds = liveFileDescriptorCount()
if (baselineFds >= 0 && finalFds >= 0) {
val fdGrowth = finalFds - baselineFds
assertTrue(
fdGrowth <= 16,
"FD count grew by $fdGrowth across $sessions sessions " +
"(baseline=$baselineFds final=$finalFds). UDP socket / pipe leak.",
)
}
}
/** True if a UDP send on the socket throws — i.e. close() has run. */
@@ -217,4 +236,18 @@ class QuicConnectionDriverLifecycleTest {
}
private fun liveThreadCount(): Int = Thread.getAllStackTraces().keys.size
/**
* Linux: count entries in `/proc/self/fd`. macOS / Windows / other
* platforms don't expose this path; return -1 so the caller
* silently skips the FD assertion. This is a strictly additive
* canary no false-positive risk on platforms that can't measure.
*/
private fun liveFileDescriptorCount(): Int =
try {
val procFd = File("/proc/self/fd")
if (procFd.isDirectory) procFd.list()?.size ?: -1 else -1
} catch (_: Throwable) {
-1
}
}
@@ -0,0 +1,261 @@
/*
* Copyright (c) 2025 Vitor Pamplona
*
* Permission is hereby granted, free of charge, to any person obtaining a copy of
* this software and associated documentation files (the "Software"), to deal in
* the Software without restriction, including without limitation the rights to use,
* copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the
* Software, and to permit persons to whom the Software is furnished to do so,
* subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
* FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
* COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
* AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
* WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
package com.vitorpamplona.quic.connection
import com.vitorpamplona.quic.frame.StreamFrame
import com.vitorpamplona.quic.stream.StreamId
import com.vitorpamplona.quic.tls.InProcessTlsServer
import com.vitorpamplona.quic.tls.PermissiveCertificateValidator
import kotlinx.coroutines.flow.toList
import kotlinx.coroutines.runBlocking
import kotlin.test.Test
import kotlin.test.assertEquals
import kotlin.test.assertTrue
/**
* Long-form heap-sampling canary for soak target #1 (memory growth).
*
* **DEFAULT-SKIPPED.** This test is gated by the `quicSoakSeconds`
* system property, propagated from the `-PquicSoakSeconds=N` Gradle
* property by `quic/build.gradle.kts`. Without the property the
* test method early-returns with a printed "SKIPPED" line, so a
* plain `./gradlew test` run keeps it out of CI's critical path
* (~40 ms overhead from the runBlocking + property check).
*
* To run the production-shaped 30-minute soak from the audio-rooms
* prompt:
*
* ./gradlew :quic:jvmTest \
* --tests 'com.vitorpamplona.quic.connection.QuicHeapSoakTest' \
* -PquicSoakSeconds=1800
*
* Quick local sanity (30 seconds, ~25 K stream lifecycles):
*
* ./gradlew :quic:jvmTest \
* --tests 'com.vitorpamplona.quic.connection.QuicHeapSoakTest' \
* -PquicSoakSeconds=30
*
* Test shape: drive moq-lite-shaped peer-uni stream churn through an
* [InMemoryQuicPipe] at a fixed rate. Sample
* `Runtime.getRuntime().totalMemory() - freeMemory()` at six evenly-
* spaced points across the run, calling `System.gc()` first to
* minimise allocator-noise. The post-warmup baseline is the second
* sample; the acceptance criterion (10 MB max growth past warmup,
* direct from the soak prompt) is checked against the final sample.
*
* Why six samples: enough granularity to spot a slow drift while
* keeping the per-sample overhead small relative to the run. The
* second sample (~17 % into the run) is the warmup baseline the
* first sample is just before any work, so it understates ambient
* usage and would inflate apparent growth.
*/
class QuicHeapSoakTest {
@Test
fun heapStaysFlatUnderModulatedStreamChurn() =
runBlocking {
val durationSec =
System.getProperty("quicSoakSeconds")?.toIntOrNull()
if (durationSec == null || durationSec <= 0) {
// Default-skip path: `./gradlew test` runs this test as a
// no-op fast-pass. Opt in via `-PquicSoakSeconds=N` (1800
// for the 30-minute soak from the audio-rooms prompt).
// We don't use kotlin.test's assumption because the JVM
// backend differs between JUnit4 / JUnit5; an early
// return prints the reason and keeps CI fast.
println(
"[QuicHeapSoakTest] SKIPPED — set -PquicSoakSeconds=N " +
"to run (e.g. 30 for sanity, 1800 for production-shape).",
)
return@runBlocking
}
// Open the pipe; the handshake should be cheap relative to
// the soak duration.
val (client, pipe) = newConnectedClient()
// moq-lite production rate: ~50 peer-uni streams per second
// (one Opus frame per stream). A 1 800 s run mints 90 000
// streams; a 30 s scaled-down run still moves through 1 500
// generations of churn — plenty for retirement to fire
// hundreds of times.
val streamsPerSecond = 50
val totalStreams = durationSec.toLong() * streamsPerSecond
val perPayload = 32 // ~Opus frame envelope
val sampleCount = 6
val streamsPerSample = (totalStreams / sampleCount).coerceAtLeast(1L)
val samples = LongArray(sampleCount)
val streamCountAtSample = LongArray(sampleCount)
var nextStreamId = StreamId.build(StreamId.Kind.SERVER_UNI, 0L)
var streamsDelivered = 0L
var sampleIdx = 0
while (sampleIdx < sampleCount) {
val target = (sampleIdx + 1) * streamsPerSample
while (streamsDelivered < target) {
val payload = ByteArray(perPayload) { (streamsDelivered.toInt() and 0xFF).toByte() }
val packet =
pipe.buildServerApplicationDatagram(
listOf(
StreamFrame(
streamId = nextStreamId,
offset = 0L,
data = payload,
fin = true,
),
),
)!!
feedDatagram(client, packet, nowMillis = 0L)
client.streamById(nextStreamId)?.incoming?.toList()
nextStreamId += 4L
streamsDelivered += 1
// Periodically drain so retireFullyDoneStreamsLocked
// runs at moq-lite-realistic intervals (per ~50
// streams). Without periodic drains the working set
// grows to streamsPerSample which is misleading vs
// the production rate.
if (streamsDelivered % streamsPerSecond == 0L) {
drainAll(client, pipe)
}
}
drainAll(client, pipe)
samples[sampleIdx] = sampleHeapBytes()
streamCountAtSample[sampleIdx] = client.streamsListLocked().size.toLong()
sampleIdx += 1
}
// Sanity: every stream we *actually* minted (streamsDelivered;
// streamsPerSample integer-truncates so it can be < totalStreams)
// must have been retired by the last sample. drainAll between
// batches drives retirement. If retirement regressed, the gap
// here would surface immediately.
assertEquals(
streamsDelivered,
client.retiredStreamsCount,
"every minted peer-uni stream must have been retired by end of soak",
)
// Working set must have stayed bounded throughout — never
// larger than the per-sample batch size (drainAll runs at
// every streamsPerSecond boundary, so the steady-state
// bound is ≤ streamsPerSecond entries).
val maxWorkingSet = streamCountAtSample.max()
assertTrue(
maxWorkingSet <= streamsPerSecond.toLong(),
"tracker working set must stay ≤ $streamsPerSecond entries; observed max=$maxWorkingSet " +
"(samples=${streamCountAtSample.toList()})",
)
// Heap acceptance — direct from the audio-rooms prompt:
// "no monotonic growth past handshake-stable (10 MB)".
// Use sample 1 (post-warmup) as baseline, sample 5 (final)
// as the steady-state probe.
val warmup = samples[1]
val finalSample = samples[sampleCount - 1]
val growthMb = (finalSample - warmup).toDouble() / (1024.0 * 1024.0)
// Print the sample profile so a CI failure has actionable
// forensic data (which sample first crossed the threshold,
// roughly when in the run).
val profile =
samples.indices.joinToString(prefix = "[", postfix = "]") { i ->
val mb = samples[i].toDouble() / (1024.0 * 1024.0)
"%.1fMB".format(mb)
}
System.err.println(
"[QuicHeapSoakTest] duration=${durationSec}s totalStreams=$totalStreams " +
"samples=$profile retired=${client.retiredStreamsCount}",
)
assertTrue(
growthMb <= 10.0,
"heap grew by %.1f MB past warmup (sample 1 = %.1f MB → final = %.1f MB); ".format(
growthMb,
warmup.toDouble() / 1024.0 / 1024.0,
finalSample.toDouble() / 1024.0 / 1024.0,
) + "samples=$profile. Acceptance: ≤ 10 MB.",
)
}
private fun sampleHeapBytes(): Long {
// Best-effort GC: a single System.gc() is a hint, not a
// guarantee. Three passes with a small sleep between gives
// G1/CMS/ZGC enough time to actually settle. (We avoid
// `System.runFinalization()` — deprecated since Java 18 and
// unreliable on modern collectors.)
repeat(3) {
System.gc()
Thread.sleep(20)
}
val rt = Runtime.getRuntime()
return rt.totalMemory() - rt.freeMemory()
}
private fun drainAll(
client: QuicConnection,
pipe: InMemoryQuicPipe,
) {
while (true) {
val out = drainOutbound(client, nowMillis = 0L) ?: break
pipe.decryptClientApplicationFrames(out)
}
}
private fun newConnectedClient(): Pair<QuicConnection, InMemoryQuicPipe> =
runBlocking {
val client =
QuicConnection(
serverName = "soak.test",
config =
QuicConnectionConfig(
initialMaxStreamsBidi = 4096,
initialMaxStreamsUni = 65_536,
initialMaxData = 16L * 1024 * 1024,
initialMaxStreamDataBidiLocal = 64L * 1024,
initialMaxStreamDataBidiRemote = 64L * 1024,
initialMaxStreamDataUni = 64L * 1024,
),
tlsCertificateValidator = PermissiveCertificateValidator(),
)
val serverScid = ConnectionId.random(8)
val tlsServer =
InProcessTlsServer(
transportParameters =
TransportParameters(
initialMaxData = 16L * 1024 * 1024,
initialMaxStreamDataBidiLocal = 64L * 1024,
initialMaxStreamDataBidiRemote = 64L * 1024,
initialMaxStreamDataUni = 64L * 1024,
initialMaxStreamsBidi = 4096,
initialMaxStreamsUni = 65_536,
initialSourceConnectionId = serverScid.bytes,
originalDestinationConnectionId = client.destinationConnectionId.bytes,
).encode(),
)
val pipe =
InMemoryQuicPipe(
client = client,
initialDcid = client.destinationConnectionId.bytes,
serverScid = serverScid,
tlsServer = tlsServer,
)
client.start()
pipe.drive(maxRounds = 16)
assertEquals(QuicConnection.Status.CONNECTED, client.status)
client to pipe
}
}